LLM API
Text, vision, speech, video and OCR behind one OpenAI-compatible key, with uptime monitoring, spend controls and an invoice instead of credits.
Read the profileRunner-up: Google Vertex AI
OpenRouter alternatives · updated 21 September 2026
OpenRouter got you started. Here is where to route next.
We compared 20 routers, gateways and inference clouds on cost, control, latency and compliance. Answer ten questions and get a shortlist built for your stack — not a generic top ten.
20alternatives compared
6questions in the planner
10criteria applied to each
0paid placements
Why teams look elsewhere
Four reasons people outgrow a single router
~5%
A fee on every token
A platform margin on top of provider prices is invisible at prototype volume and a real line item at production volume. Routers that bill provider rates, or let you bring your own keys, close that gap.
EU
Prompts that cannot travel
Once legal reviews the data path, a US-only proxy in front of your prompts stops being acceptable. Regional hosting, self-hosted gateways and cloud platforms keep traffic inside a perimeter you can name.
p95
Latency budgets
Voice agents and live UX spend their whole budget in the first second. A direct inference provider with no extra hop usually beats any router on speed, whatever the catalogue says.
360°
Control and visibility
Per-team budgets, spend caps, traces, caching and one invoice stop being nice-to-have the moment a second team starts calling models on the company card.
When to stay on OpenRouter: you are prototyping, you want one bill across hundreds of models, and a small fee is cheaper than running infrastructure. The planner says so too when that is the honest answer.
What the gateway itself costs you
Token prices are roughly the same wherever you route. The difference is the fee sitting on top. Drag the slider to your monthly spend.
$1,000 / month$12,000 a year in tokens
$100$1k$10k$100k
Gateway fee only — the tokens themselves are billed on top and cost about the same everywhere. Percentage fees scale with you; a flat fee stops growing. Self-hosting costs nothing in fees but you pay for servers and the engineer who keeps them up. Fee shapes taken from each vendor's own pricing page, checked 21 September 2026 — the full working is in our comparison write-up.
Route planner
Ten questions to your shortlist
The live board re-ranks every alternative as you answer. No email, no gate, and you can share the result with the link in your address bar.
Question 1 of 10
Editor's picks
Best for mixed modalities
Best self-hosted
Best drop-in swap
Lowest latency
Cheapest open models
Best for enterprise
The short answer, by need
If you already know which wall you hit, start here.
LiteLLM
An open-source proxy for 100+ providers with virtual keys, budgets and fallbacks — the closest thing to OpenRouter you can run inside your own perimeter.
Read the profileRunner-up: Kong AI Gateway
Requesty
The same one-key, one-bill shape with cost-aware routing and caching on top. Migration is a base URL and a key, nothing else.
Read the profileRunner-up: Unify
Groq
Purpose-built inference hardware for open-weight models. The pick when voice agents or real-time UX own the latency budget.
Read the profileRunner-up: Fireworks AI
DeepInfra
Aggressive per-token pricing on Llama, Qwen, DeepSeek and friends, on an OpenAI-compatible API with no platform fee layered on top.
Read the profileRunner-up: Together AI
AWS Bedrock
Frontier and open models under the AWS contract, IAM boundary and compliance paperwork your procurement team already signed off.
Read the profileRunner-up: Azure AI Foundry
Every alternative, side by side
The same data the planner runs on. Filter by what you need, or click a column heading to sort.
| Alternative | Model access | Modalities | Hosting | Routing & failover | Pricing shape |
|---|---|---|---|---|---|
| 400+ text models | Text, some vision | Managed cloud | Automatic provider fallback | Provider price + platform fee, prepaid credits | |
| 400+ models, 30+ providers | Text, vision, speech, video, OCR | Managed cloud, EU region | Failover + live uptime monitoring | Pay per token with volume discounts, invoiced | |
| 200+ open models | Text, image, embeddings | Managed cloud | Single provider, no cross-vendor failover | Per token, plus hourly dedicated GPUs | |
| 100+ open models | Text, vision, audio, image | Managed cloud | Single provider | Per token, plus on-demand GPU hours | |
| ~20 curated models | Text, speech-to-text | Managed cloud | Single provider | Per token, free developer tier | |
| 100+ open models | Text, image, speech, embeddings | Managed cloud | Single provider | Per token, pay as you go | |
| 200+ models | Text, image, video, speech | Managed cloud | Single provider | Per token or per generation | |
| Thousands of community models | Image, video, audio, text | Managed cloud | None | Per second of compute | |
| AWS Bedrock |
~50 models from selected vendors | Text, image, embeddings | Managed cloud, many regions incl. EU | Cross-region inference, no cross-vendor routing | Per token, on your AWS bill; provisioned throughput available |
| OpenAI + 1000s of catalogue models | Text, image, speech, vision | Managed cloud, EU data boundary | Regional deployments, no cross-vendor routing | Per token or provisioned units, on the Azure bill | |
| Gemini + model garden | Text, image, video, speech | Managed cloud, EU regions | Regional endpoints, no cross-vendor routing | Per token or per media unit, on the GCP bill | |
| Bring your own provider keys | Whatever your providers support | Managed cloud or self-hosted | Conditional routing, retries, fallbacks | Per request tiers, free developer plan, enterprise self-host | |
| 100+ providers via your own keys | Whatever your providers support | Self-hosted (managed tier available) | Load balancing, retries, fallbacks | Free and open source, paid enterprise tier | |
| Bring your own provider keys | Whatever your providers support | Self-hosted or hybrid | Policy-based routing and rate limits | Open source core, enterprise licence | |
| Bring your own keys + Workers AI | Text, image, speech via Workers AI | Managed edge, EU options | Fallbacks, caching, rate limits | Free tier, usage-based extras | |
| Your own model deployments | Whatever you deploy | Your cloud account / VPC | You build it | Platform fee on top of your own compute | |
| Open models via partner providers | Text, image, audio, embeddings | Managed cloud | Provider selection per model | Free credits, then pass-through provider pricing | |
| 150+ models | Text, vision | Managed cloud | Cost-aware routing and fallbacks | Usage fee on top of provider prices | |
| 100+ AI services across vendors | Text, OCR, speech, image, translation | Managed cloud, EU company | Provider fallback per task | Per call, with a monthly plan | |
| Major providers and open models | Text | Managed cloud | Live benchmark-based routing | Provider price plus routing fee |
Facts checked against each vendor's own documentation and pricing pages on 21 September 2026. Model counts move weekly — treat them as orders of magnitude, not guarantees.
Coverage at a glance
Every alternative against the twelve things the planner scores on. Filled means the platform covers it as a first-class capability; empty means it does not — no half-credit for "on the roadmap".
| Alternative | Text | Vision | Speech | Video | Managed cloud | EU hosting | Self-host | Failover | Unified billing | White label | Enterprise | Drop-in swap |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| covered | covered | covered | covered | covered | covered | not covered | covered | covered | covered | covered | covered | |
| covered | covered | not covered | not covered | covered | not covered | not covered | not covered | not covered | not covered | not covered | covered | |
| covered | not covered | not covered | not covered | not covered | covered | covered | covered | not covered | not covered | covered | not covered | |
| AWS Bedrock | covered | covered | not covered | not covered | covered | covered | not covered | covered | covered | not covered | covered | not covered |
| covered | covered | covered | not covered | covered | covered | not covered | covered | covered | not covered | covered | not covered | |
| covered | covered | covered | not covered | covered | covered | not covered | covered | not covered | not covered | not covered | covered | |
| covered | covered | covered | not covered | covered | not covered | not covered | not covered | not covered | not covered | not covered | covered | |
| covered | covered | covered | not covered | covered | covered | not covered | not covered | covered | not covered | not covered | covered | |
| covered | covered | covered | not covered | covered | not covered | not covered | covered | not covered | not covered | not covered | covered | |
| covered | covered | covered | covered | covered | covered | not covered | covered | covered | not covered | covered | not covered | |
| covered | not covered | covered | not covered | covered | not covered | not covered | not covered | not covered | not covered | not covered | covered | |
| covered | covered | covered | not covered | covered | not covered | not covered | not covered | not covered | not covered | not covered | covered | |
| covered | not covered | not covered | not covered | not covered | covered | covered | covered | covered | not covered | covered | not covered | |
| covered | covered | covered | not covered | not covered | covered | covered | not covered | covered | not covered | not covered | not covered | |
| covered | covered | covered | covered | covered | not covered | not covered | not covered | not covered | not covered | not covered | covered | |
| covered | covered | covered | not covered | covered | covered | covered | covered | covered | not covered | covered | covered | |
| not covered | covered | covered | covered | covered | not covered | not covered | not covered | not covered | not covered | not covered | not covered | |
| covered | covered | not covered | not covered | covered | not covered | not covered | not covered | covered | not covered | not covered | covered | |
| covered | covered | not covered | not covered | covered | not covered | not covered | not covered | not covered | not covered | not covered | covered | |
| covered | not covered | not covered | not covered | covered | not covered | not covered | covered | not covered | not covered | not covered | covered |
CoveredNot coveredSame dataset the planner and the table above run on.
The 20 profiles
What each one actually is, who it suits, and the part their marketing page leaves out.
OpenRouter
A marketplace-style router in front of hundreds of hosted models, with a single OpenAI-compatible endpoint, credits-based billing and automatic fallbacks between upstream providers.
- Trying many models quickly, side projects, teams that want the widest text-model catalogue without contracts.
- A per-request fee sits on top of provider prices, support is community-first, and speech, video and OCR coverage is thin compared with the text catalogue.
- Provider price + platform fee, prepaid credits
LLM API
A managed router covering 400+ models from 30+ providers behind one key, with per-model uptime monitoring, spend controls, invoices instead of credits, and a white-label router other companies resell under their own brand.
- Teams running mixed modalities in production, agencies and platforms that need one invoice, and companies reselling AI access.
- No self-hosted edition — the gateway runs as a managed service.
- Pay per token with volume discounts, invoiced
Together AI
An inference cloud running open-weight models on its own GPU fleet, plus fine-tuning and dedicated capacity for steady workloads.
- Open-model workloads at volume where per-token price and throughput matter more than closed-model access.
- No access to closed frontier models from OpenAI or Anthropic through the same endpoint.
- Per token, plus hourly dedicated GPUs
Fireworks AI
A performance-focused inference platform for open-weight models, with speculative decoding, LoRA tuning and on-demand deployments.
- Latency-sensitive production apps built on open models.
- Catalogue is narrower than a marketplace router and closed frontier models are not included.
- Per token, plus on-demand GPU hours
Groq
Custom LPU hardware serving a curated set of open models at very high throughput through an OpenAI-compatible API.
- Voice agents, streaming chat and anything where response speed is the product.
- Small catalogue, capacity limits at peak, and no cross-provider routing.
- Per token, free developer tier
DeepInfra
A low-cost inference provider with an OpenAI-compatible endpoint across open text, embedding, image and speech models.
- Cost-sensitive batch and background workloads.
- Fewer enterprise controls, lighter SLAs, no closed frontier models.
- Per token, pay as you go
Novita AI
An inference cloud spanning open LLMs plus image, video and speech generation, also available as GPU instances.
- Media-heavy products that also need text models at low cost.
- Quality of service varies by model; enterprise paperwork is lighter than a hyperscaler's.
- Per token or per generation
Replicate
A hosting layer for thousands of community-published models, including image, video and audio pipelines, with custom model deployment.
- Generative media, experimentation and running your own containerised model.
- Cold starts, per-second billing that is hard to forecast, and inconsistent model maintenance.
- Per second of compute
AWS Bedrock
Amazon's managed model service with Anthropic, Meta, Mistral, Amazon and other families, IAM controls, VPC access and regional isolation.
- Enterprises already on AWS that need procurement, data residency and compliance more than catalogue breadth.
- Model availability differs by region, quotas need raising, and the developer experience is heavier than a router.
- Per token, on your AWS bill; provisioned throughput available
Azure AI Foundry
Microsoft's AI platform hosting OpenAI models alongside open and partner models, with private networking, content filters and EU data boundary options.
- Microsoft-committed enterprises needing OpenAI models with corporate compliance.
- Deployment quotas per region, and the catalogue outside OpenAI is uneven.
- Per token or provisioned units, on the Azure bill
Google Vertex AI
Google's managed platform for Gemini plus a model garden of partner and open models, with grounding, tuning and regional endpoints.
- Teams standardising on Gemini, or needing long-context and multimodal models under GCP contracts.
- Authentication and project setup are heavier than an API key; pricing differs per modality.
- Per token or per media unit, on the GCP bill
Portkey
A gateway that sits between your app and any provider key you bring, adding routing rules, caching, retries, budgets, logs and guardrails. Available managed or self-hosted.
- Platform teams that already hold provider contracts and need control, tracing and spend limits.
- You still buy the models elsewhere — it is a control plane, not a catalogue.
- Per request tiers, free developer plan, enterprise self-host
LiteLLM
An MIT-licensed Python SDK and proxy that normalises 100+ provider APIs into the OpenAI format, with virtual keys, budgets and logging. Run it in your own cluster.
- Engineering teams that want full control, no vendor in the data path, and provider prices with no markup.
- You own uptime, upgrades and scaling; the enterprise features sit behind a paid tier.
- Free and open source, paid enterprise tier
Kong AI Gateway
Kong's gateway extended with AI plugins for multi-provider routing, prompt guards, token rate limiting and semantic caching, deployable on-prem.
- Enterprises that already standardise API traffic on Kong and want AI under the same policy layer.
- Infrastructure-team territory: no model catalogue, no consumer-grade onboarding.
- Open source core, enterprise licence
Cloudflare AI Gateway
A thin edge layer that proxies calls to providers, adding analytics, caching, retries and rate limits, with Workers AI for models hosted by Cloudflare.
- Teams on Cloudflare who want cheap observability and caching without changing providers.
- Not a catalogue or a billing layer; governance features are lighter than dedicated gateways.
- Free tier, usage-based extras
Anyscale
A managed Ray platform for serving and fine-tuning models on your own infrastructure, including deployment into your VPC.
- ML teams with custom models and existing cloud commitments.
- Heaviest setup on this list; no ready-made catalogue or router.
- Platform fee on top of your own compute
Hugging Face Inference
Inference providers on the Hub route calls to partners such as Together, Fireworks and Novita behind one token, alongside dedicated endpoints.
- Open-model research, prototypes and anything already anchored on the Hub.
- No closed frontier models, and production guarantees depend on the upstream partner.
- Free credits, then pass-through provider pricing
Requesty
A multi-provider routing layer with caching, fallbacks, per-user budgets and analytics aimed at cutting model spend.
- Teams whose main pain is an unpredictable model bill.
- Younger platform with a smaller catalogue and shorter operating history.
- Usage fee on top of provider prices
Eden AI
An aggregator covering OCR, translation, speech, moderation and generative models from many vendors under a single key and invoice, with an EU base.
- Products that need document, speech and translation AI alongside LLMs.
- Task-oriented rather than model-oriented; per-call pricing can exceed direct provider rates.
- Per call, with a monthly plan
Unify
A router that scores providers on live latency, throughput and cost, then sends each request to the endpoint that best matches your constraints.
- Engineers who want routing decisions made on measurements rather than a fixed preference list.
- Smaller catalogue and community; best suited to text workloads.
- Provider price plus routing fee
Short version, by situation
If you scrolled past the planner, start here.
You are prototyping and want maximum choice
Stay on OpenRouter, or use Hugging Face Inference. The per-request fee is irrelevant at small volume and the catalogue is the widest there is.
You are in production across text, voice and video
LLM API covers all four modalities behind one key with uptime monitoring and a single invoice. Novita AI is the budget alternative if you can accept lighter guarantees.
Your security team wants nothing in the data path
LiteLLM self-hosted, or Kong AI Gateway if you already run Kong. You keep provider prices with no markup and own the uptime.
Procurement has to sign off
AWS Bedrock, Azure AI Foundry or Google Vertex AI — the model spend lands on a cloud contract you already have, with residency and compliance attached.
How we compared them
Ten criteria, applied the same way to every entry, from public documentation and hands-on use.
- Model breadth and how fast new releases land
- Modality coverage: text, vision, speech, video, OCR
- Routing, retries and failover behaviour
- Pricing shape: markup, credits, invoices, commitments
- Spend controls, budgets and per-team reporting
- Data residency and regional hosting
- Self-hosting and the licence that comes with it
- White-label and reseller support
- Enterprise readiness: SLAs, SOC 2, procurement
- Migration effort from an OpenAI-compatible endpoint
What we could not measure
Latency and uptime vary by region, model and hour, so we did not publish a single speed ranking — any number we quoted would be stale by the time you read it. Support quality is reported, not tested. Where a vendor does not publish a fact, the table says so instead of guessing.
Read the full methodology write-up
FAQ
Questions we get a lot
Still unsure after the planner? These come up in almost every migration call.
What is the best free OpenRouter alternative?
For self-hosting, LiteLLM is open source and free to run — you pay only your upstream providers. For hosted, Cloudflare AI Gateway's core features are free and sit in front of your own provider keys.
Can I switch without rewriting code?
Usually yes. Most alternatives on this page expose an OpenAI-compatible endpoint, so you change the base URL, the API key and sometimes the model-ID format. The table's pricing and routing columns flag the ones that need more than that.
Gateway or inference provider — what is the difference?
A gateway routes your requests to providers you already pay and adds control: budgets, caching, logs, failover. An inference provider runs the models itself. Plenty of teams use both, a gateway in front and two or three providers behind it.
What does BYOK mean?
Bring Your Own Key: you plug in your own OpenAI, Anthropic or Google keys, so those providers bill you directly and the router charges only for its own layer — or for nothing at all.
When should I just stay on OpenRouter?
When you are prototyping, want the widest text catalogue with no contract, and the per-request fee is smaller than the cost of running anything yourself. The planner says so too when that is the honest answer.
Is anyone paying to be on this list?
No. LLM API is our own product and is labelled as such wherever it appears. Every other entry is here on merit, scored on the same criteria from public documentation and hands-on use.
Try the one built for mixed workloads
LLM API gives you 400+ text, vision, speech and video models behind a single OpenAI-compatible key, with uptime monitoring, spend controls and one invoice. Migration is a base URL and a key.
Start free